Nn growth apparatus, information processing apparatus, method for producing neural network information, and program

Abstract

An NN growth apparatus includes: a start point storage in which pieces of firing start point information are stored, each piece of firing start point information containing an information identifier of feature information of image information and node identifiers; a goal storage in which pieces of goal information are stored, the pieces of goal information specifying goals respectively corresponding to states; a state determination unit that determines one state, using the sound information; a feature acquisition unit that acquires feature information from the image information; a firing node determination unit that determines node identifiers corresponding to the pieces of feature information from the start point storage unit, and determines a node identifier of a firing node connected to the nodes; and a growth unit that acquires the goal information paired with the one state, and performs processing to grow node information or edge information corresponding to node identifiers.

1 . An NN growth apparatus comprising:
 an NN storage unit in which neural network information is stored, the neural network information containing two or more pieces of node information each containing a node identifier, and one or more pieces of edge information each containing an edge identifier and specifying a connection between nodes;   a start point storage unit in which one or more pieces of firing start point information are stored, each piece of firing start point information containing an information identifier that identifies feature information of image information and one or more node identifiers that identify nodes that initially fire when the feature information is accepted;   a goal storage unit in which pieces of goal information are stored, the pieces of goal information specifying goals respectively corresponding to two or more states that include positive and negative states;   an information acceptance unit that accepts image information and sound information;   a state determination unit that determines one state from among the two or more states, using the sound information accepted by the information acceptance unit;   a feature acquisition unit that acquires, using the image information accepted by the information acceptance unit, one or more pieces of feature information corresponding to the image information;   a firing node determination unit that determines one or more node identifiers of one or more firing nodes corresponding to the one or more pieces of feature information acquired by the feature acquisition unit, from the start point storage unit, and determines a node identifier of a firing node that is connected to the one or more nodes identified by the one or more node identifiers by edges and to which the feature information is to be transferred; and   a growth unit that acquires the goal information paired with the one state determined by the state determination unit, and, using the goal information, performs processing to grow node information or edge information corresponding to one or more node identifiers of the one or more node identifiers determined by the firing node determination unit.   
     
     
         2 . The NN growth apparatus according to  claim 1 , further comprising:
 a window determination unit that sequentially determines, from a still image contained in the image information accepted by the information acceptance unit, a sliding window, which is a partial region of the still image,   wherein the feature acquisition unit uses a partial image corresponding to the sliding window sequentially determined by the window determination unit, to acquire one or more pieces of feature information corresponding to the partial image.   
     
     
         3 . The NN growth apparatus according to  claim 1 ,
 wherein the feature information contains an information identifier that identifies information and an information amount that indicates a size of the information, and   the firing node determination unit judges whether or not one or more pieces of feature information transferred from one or more other nodes connected by edges satisfy one or more firing conditions related to the feature information, and determines the node identifier of a node judged to satisfy the firing conditions.   
     
     
         4 . The NN growth apparatus according to  claim 1 ,
 wherein the node information contains node position information that specifies a position of the node,   the goal information contains goal position information that specifies a position of the goal, or goal direction information that indicates a direction of the goal, and   the growth unit performs edge generation processing to generate and accumulate edge information for an edge extending from one or more nodes identified by one or more node identifiers among the one or more node identifiers determined by the firing node determination unit in the direction indicated by the goal information that is paired with the one state determined by the state determination unit.   
     
     
         5 . The NN growth apparatus according to  claim 4 ,
 wherein the firing node determination unit accumulates count information related to a count of the determined node identifiers in association with the node identifiers, and   the growth unit performs the edge generation processing on a node identified by a node identifier corresponding to the count information that matches an edge generation condition.   
     
     
         6 . The NN growth apparatus according to  claim 1 ,
 wherein the node information contains node position information that specifies a position of the node,   the goal information contains goal position information that specifies a position of the goal, or goal direction information that indicates a direction of the goal, and   the growth unit performs edge growth processing to acquire and accumulate edge information in which an edge extending from one or more nodes identified by one or more node identifiers among the one or more node identifiers determined by the firing node determination unit is grown in a direction indicated by the goal information paired with the one state determined by the state determination unit.   
     
     
         7 . The NN growth apparatus according to  claim 6 ,
 wherein the firing node determination unit accumulates count information related to a count of the determined node identifiers in association with the node identifiers, and   the growth unit performs the edge growth processing on a node identified by a node identifier corresponding to the count information that matches an edge generation condition.   
     
     
         8 . The NN growth apparatus according to  claim 1 ,
 wherein each node is a soma,   the edges include an axon and a dendrite, and   the edge information contains axon information that contains an axon identifier and axon position information indicating a position of the axon, and dendrite information that contains a dendrite identifier and dendrite position information indicating a position of the dendrite.   
     
     
         9 . An information processing apparatus comprising:
 an NN storage unit in which neural network information accumulated by the NN growth apparatus according to  claim 1  is stored;   an information acceptance unit that accepts acceptance information that is at least one of image information and sound information;   a feature acquisition unit that acquires one or more pieces of feature information corresponding to the acceptance information accepted by the information acceptance unit;   an information transfer unit that determines one or more node identifiers of one or more firing nodes, corresponding to the one or more pieces of feature information acquired by the feature acquisition unit, from a start point storage unit in which one or more pieces of firing start point information are stored, each piece of firing start point information containing an information identifier that identifies feature information of acceptance information and one or more node identifiers that identify nodes that initially fire when the feature information is accepted, and determines a node identifier of a firing node that is connected to the one or more nodes identified by the one or more node identifiers by edges and to which the feature information is to be transferred;   a firing pattern acquisition unit that acquires a firing pattern using the one or more node identifiers determined by the information transfer unit;   an output information acquisition unit that acquires output information corresponding to the firing pattern acquired by the firing pattern acquisition unit; and   an information output unit that outputs the output information.   
     
     
         10 . The information processing apparatus according to  claim 8 , further comprising:
 a temperature acceptance unit that accepts temperature information,   wherein the information transfer unit performs information transfer processing to transfer, from a node that has fired, the feature information corresponding to the node that has fired, to a node that is to fire next, and changes a processing time for the information transfer processing based on the temperature information accepted by the temperature acceptance unit.   
     
     
         11 . A method for producing neural network information realized by: an NN storage unit in which neural network information is stored, the neural network information containing two or more pieces of node information each containing a node identifier, and one or more pieces of edge information each containing an edge identifier and specifying a connection between nodes; a start point storage unit in which one or more pieces of firing start point information are stored, each piece of firing start point information containing an information identifier that identifies feature information of image information and one or more node identifiers that identify nodes that initially fire when the feature information is accepted; a goal storage unit in which pieces of goal information are stored, the pieces of goal information specifying goals respectively corresponding to two or more states that include positive and negative states; an information acceptance unit, a state determination unit, a feature acquisition unit, a firing node determination unit, and a growth unit, the method comprising:
 an information acceptance step in which the information acceptance unit accepts image information and sound information; 
 a state determination step in which the state determination unit determines one state from among the two or more states, using the sound information accepted in the information acceptance step; 
 a feature acquisition step in which the feature acquisition unit acquires, using the image information accepted in the information acceptance step, one or more pieces of feature information corresponding to the image information; 
 a firing node determination step in which the firing node determination unit determines one or more node identifiers of one or more firing nodes corresponding to the one or more pieces of feature information acquired in the feature acquisition step, from the start point storage unit, and determines a node identifier of a firing node that is connected to the one or more nodes identified by the one or more node identifiers by edges and to which the feature information is to be transferred; and 
 a growth step in which the growth unit acquires the goal information paired with the one state determined by the state determination unit, and, using the goal information, performs processing to grow node information or edge information corresponding to one or more node identifiers of the one or more node identifiers determined by the firing node determination unit. 
 
     
     
         12 . A program that enables a computer that can access: an NN storage unit in which neural network information is stored, the neural network information containing two or more pieces of node information each containing a node identifier, and one or more pieces of edge information each containing an edge identifier and specifying a connection between nodes; a start point storage unit in which one or more pieces of firing start point information are stored, each piece of firing start point information containing an information identifier that identifies feature information of image information and one or more node identifiers that identify nodes that initially fire when the feature information is accepted; and a goal storage unit in which pieces of goal information are stored, the pieces of goal information specifying goals respectively corresponding to two or more states that include positive and negative states, to function as:
 an information acceptance unit that accepts image information and sound information;   a state determination unit that determines one state from among the two or more states, using the sound information accepted by the information acceptance unit;   a feature acquisition unit that acquires, using the image information accepted by the information acceptance unit, one or more pieces of feature information corresponding to the image information;   a firing node determination unit that determines one or more node identifiers of one or more firing nodes corresponding to the one or more pieces of feature information acquired by the feature acquisition unit, from the start point storage unit, and determines a node identifier of a firing node that is connected to the one or more nodes identified by the one or more node identifiers by edges and to which the feature information is to be transferred; and   a growth unit that acquires the goal information paired with the one state determined by the state determination unit, and, using the goal information, performs processing to grow node information or edge information corresponding to one or more node identifiers of the one or more node identifiers determined by the firing node determination unit.